An End-to-End Segmentation Network for Unregistered muti-modality magnetic resonance images

Background: Medical image segmentation holds a crucial position in facilitating medical diagnosis. Multi-modal segmentation tasks have been widely used in the field of medical diagnosis. At present, in multi-modal segmentation tasks based on deep learning, almost all methods require different modalities to have the same size and voxel to meet the requirements of network input. Therefore, the general methods require pre-processing for registration of different modalities, which is a complex and time-consuming process. Methods: In this work, we introduce an innovative framework specifically designed for segmenting unregistered multi-modal magnetic resonance images (MRI). This method builds upon the 3D-UNet architecture, inte-grating registration into the segmentation network for enhanced performance. Results: we validate our model on the public WMH Segmentation Challenge dataset. The experimental findings reveal that our proposed method surpasses the conventional 3D U-Net in terms of the average Dice coefficient. In short, the proposed method can avoid complex registration pretreatment processes while obtaining a good segmentation result.